Hacker News Reader

One example: Bad weather/road conditions where human might have problems but autonomic cars currently will go "no, you drive it".

This. Once autonomic cars can safely and with confidence steer through a crowded IKEA parking lot on Saturday morning when it's raining we can say they are ready.

Even if they can do that, they cannot drive in most of Canada almost half the year -- we have snow and icy roads 3-6 months a year.

Since human drivers can do this, what's missing from autonomous systems to perform here as well? More training data? Different optical or radar systems? Assuming our ability to drive isn't some God given gift, there shouldn't be a hard limit on what autonomous cars can do, yes?

I am sure they will do this eventually, but it is next level.

They would likely have to setup in a Northern location and run their cars in that location for a while. Also need to update their simulations to account for snow and icy roads. So more complex driving models and more challenging road detection.

Sometimes there is light snow over the whole road and there are no markings. You just have to position yourself relative to street signs or ditches where you think your lane is.

Also we then get to some more interesting challenges. Certain snow conditions can have rather bad traction like very muddy roads do. And self-driving should be able to navigate these better than regular drivers.

How can the car determine the road condition, especially if it varies (snow here vs ice under the snow in this spot vs that's a drift)? It seems like a lot of this is based on driver experience whereas traction controls today are based on things like differential wheel speed comparison. It won't determine a bad situation until you're in it.

Autonomous vehicles are good at making routine repeatable decisions based on clear data. I'm not sure how we can sense and input this sort of data clearly enough.

Eventually these autonomous vehicles should share driving condition data in real-time-ish. If one car finds a stretch of the road slippery, it should basically mark it on the map as a bad location, and then other cars and be careful.

I think weather conditions can also predict road conditions pretty well. And the combine that with historical data on where things were slippery, I bet you can create a pretty good model of what conditions to expect across a driving environment.

Although in Canada, sometimes all roads in a city are just horrid - dozens of cars in ditches or hitting each other because no one can stop. And maybe in those cases, autonomous vehicles should refuse to drive until it is improved.

True, historical data is basically what humans use too. I just wonder if there's a way to prevent the accidents that would generate that data

Where I live, the sharing is done at a meta level. I live in the Sierra Nevada mountains of California and we often get experienced snow drivers who get stuck immediately because they aren't used to driving on 25% grades that experience frequent freeze / thaw cycles (and therefore generate ice). The issues here are predictable and don't even require real-time sharing.

When someone new moves into town, someone will typically explain to them to watch for where the sun hits the ground in winter. Where that happens on a steep slope, avoid driving there when there is snow. Before they know to do this, a human will attempt to climb that hill, not make progress because of the ice, back down, and try another route. This is a rural area without much traffic, so conditions might change significantly between "real-time" updates.

So, we have human drivers "zooming out" in a couple of ways... in time/context (observing conditions that lead to certain routes becoming high risk) and in problem resolution (try another route). This is quite different from what is currently being tried algorithmically.

They don't have enough data to brute force the problem with machine learning. The dataset available will never be large enough to allow ML autonomous cars to solve every situation a human driver can. Another AI breakthrough is needed.

LLMs are so good because they have almost the entire digitized text output of humanity available for training. They may not always provide factually correct information, but it is dished out in near perfect English (or whichever language you prefer).

Personally i think if we want to recreate a human driver full AGI is needed. But, in the right conditions current autonomous cars will work well. They just need some external helpers, like being geo-fenced to a set of roads which have all the correct sign posts and markings.

> The dataset available will never be large enough to allow ML autonomous cars to solve every situation a human driver can.

Not every human driver can solve every situation that the best human driver can solve. Also, the most dangerous situations happen at high speed, which humans aren't great at handling unless we've had a lot of training. The main reason I'm optimistic about autonomous cars is that a system like Waymo's will only get better and better over the years. And newly produced cars will be just as good as "experienced" cars.

> Also, the most dangerous situations happen at high speed

This is a mantra from speed limit fans. There are a lot of dangerous situations happening at low speed (bicicle coming from left in an intersection, children playing )

If you consider those situations dangerous you aren't driving slow enough.

The other problem is that even if you have car that can do 95% of the driving and only need 5%...

...that makes the human driver, who normally would get say 10k miles of practice every year will drive... 500 miles on average.

It really does need to be "car can do all of the driving", or else it will be putting people that have zero practice in those hard situations it can't handle.

Yes, I agree. And regardless of experience, people at some point get too old to drive, but don't always want to accept that fact. Having cars that can do 100 % of the driving would be great.

You need more training data and more sensors and completely different rules that are often opposite the rules for good conditions. The things you really need for snow/ice that are hard for computers are anticipation and flexible rules.

Why is that the bar? It they are safe and useful in other situations, they can already be used in those other situations, no?

> Why is that the bar? It they are safe and useful in other situations, they can already be used in those other situations, no?

Unlike in SW, choosing a different path might lead to accidents.

My car can keep the lane very good in a sunny day, when the sin shines from above (i.e. midday).

In heavy rain or snow, i can rely only on my senses.

For an automatic driving car to be useful, it must function correctly in 100 % of the cases (ok, 99,999 %).

I was talking about general purpose autonomous driving. Special cases of autonomous driving have been in use for a while like trains (the Yurikamome in Tokyo started operation in 1995), buses driving on designated routes, Waymo cars that operated in certain areas etc.

Interesting actual useful stat would be - insurance companies covering Tesla owners, reported accident rates on that pool of owners before & after Tesla ownership.

I raise this because the pool of people buying a Tesla is different than the full driver pool. There is likely serious sampling bias in this pool of drivers due to demographics, geographics, etc..

Insurance seems really weird to me with self driving cars. It seems like the driver being insured is Tesla Motor Co not Bill Smith. Bill just owns the car, he doesn't drive it.

I've noticed a lot of automation legally still require a warm body to take the blame. Seems like eating your cake and having it too.

Ignore Tesla. Yes Elon is going to keep saying they're selling a "Self driving" car but Elon's lawyers are going to keep telling anybody who gets into an accident they were driving because Elon isn't really selling a self-driving car.

Waymo are driving, there is no warm body, in accidents where a Waymo self driving taxi struck somebody that's on Waymo, you're just a passenger the same as if you were in a New York cab, or a Uber, or if an A320 you're sat in smacks into a post because the pilots took a corner too close. You're not the one who needs insurance, because you're not driving, Waymo is.

Agree that Tesla isn't self driving in the technical sense. However, given public perception and number of units.. Tesla is what 99% to 99.9% of consumers will interact with in the "self driving" space.

Even Waymo still has some safety drivers per the article linked. And if we are going to constrain our analysis to self driving pilot programs, then we need to constrain the comparison to the same geofenced area. Arguably there are not a lot of fatal accidents per 1M miles on the local roads of Phoenix/SanFran by human drivers either.

The third important point about not comparing averages is that there are many human drivers that are outliers - people who are on a suspended license, intoxicated (over 1/2 of injured drivers are intoxicated), people who should have never had a license to begin with (either unable or needed more training) due to our joke testing, etc.

I would much rather see comparisons to safe/prudent drivers vs the average. Try to sell me autopilot as being as safe or marginally better than the average driver and I'll never use that shit - the average driver sucks.

https://www.npr.org/2022/12/13/1142598789/drivers-crashes-dr...

I would call 600% improvement over average more than "marginal"

Imagine the following scenario: your hometown is encircled by forest fires, the only way out is a single road. That road sees some abnormal conditions, smoke, vehicles on the side, maybe burning debree on the road.

Would you trust a self driving vehicle to attempt to drive your family into safety?

If not, then you probably got some food for thought about averages.

Where did you get that number? Only from the limited test scope of SF taxis? Or is this fromba more comprehensive testing?

But for these types of analyses you have to take the average because autonomous driving will improve the safety of these “average” drivers quite a bit.

I get what you’re saying though. I’ve been driving for 20 years, have had a single accident (rear ended at a stop light, not really my fault), no moving violations, and have driven about ~300,000 miles total (sounds like a lot but I was a full time lyft driver for a bit).

I’m positive I’m an outlier but I’d also trust myself a lot more than some of the current autonomous vehicles. I’d still assume with AI assistance or partial autonomy, I’d be safer on average though.

If you’re not a reckless male under the age of 25 or a persistently-intoxicated driver, it’s perfectly reasonable to want an autonomous system to be compared to YOUR safe driving profile. Reckless young men and drunks are extreme outliers of safety and they bring the average safety level way down. Being safer than average means nothing. Needs to be safer than an adult driver with clean history and cheap insurance.

Of course it means something. If every reckless drunk <25 year old was replaced by autonomous driving YOUR driving experience would automatically become safer, do you see why?

Except that’s not the proposition for early adopters. It’s the wealthier, safer driver replacing their driving with a (clearly) unsafe, inferior autonomous system. And the 18 year old drag racers continuing to drive rusted out Civics, because that’s what they can afford. So it’s a net decrease in safety in the immediate term, with the safe drivers expected to be the primary Guinea pigs in the experiment, built on a hope and dream that the tech will actually deliver eventually.

But that's not the paradigm unfolding here. The reckless drivers will shun self driving because they want to show off and have fun. The drunk driver part could be reduced after sufficient rollout, but we'll have to see how that plays out - the people whonaleays use it would use it when drinking too, but you'll still have people going "I'm fine to drive" and not using it if it's not habit.

>But that's not the paradigm unfolding here.

Yes, that is the paradigm unfolding here. That is indeed one of the many critical attractions of self driving vehicles.

>The reckless drivers will shun self driving because they want to show off and have fun.

First, this is your unsourced head canon. People with drinking problems, cognitive impairment (both temporary and permanent), etc seem likely to be a far bigger percentage.

But it doesn't matter, because they won't get the choice, which again is one of the core aspects of self driving vehicles. Right now, human driving in America is a quasi-right because it's currently intrinsically tied to "personal arbitrary time/space mechanized transportation". I say "quasi-right" because while legally it's not formally a codified right or in the Constitution or something, as a practical matter if most of the population wasn't able to drive the country would fall apart. And that is recognized by government, both executive/legislative and judicial. Losing driving rights would be a major burden for tens of millions people and would disrupt their ability to get or retain jobs, in turn causing other societal expense. So traditionally government is quite hesitant to wield that stick and almost never preemptively.

Self-driving though would decouple human driving from someone having their own car and ability to have it take them where they want, when they want like right now. That in turn would completely change to political dynamics. Driving would truly become a privilege, and getting a license could become much more like class C licenses or professional emergency responder licenses or the like are. Much, much more serious training and testing, continuing ed requirements as with other professions, and much easier to lose. Because unlike right now, losing one's drivers license won't mean losing the ability to use a car. Anyone who acts recklessly at all, even if there is no damage of any kind, can just preemptively have their license rescinded until at a minimum remediation is completed. Because driving will just be a hobby.

I read this as kind of goalpost-movey. If we can, with self driving cars, clip some of those outliers, we will have outsize impact on the total safety picture. Average skill is what we deal with all the time (On average. ;) so better than that ought to be sufficient.

That long tail of outlier clipping works on an individual human basis, too. How many accidents happen because an individual careful driver is, momentaritly, inanttentive?

"If we can, with self driving cars, clip some of those outliers, we will have outsize impact on the total safety picture."

It depends. I doubt it because the distribution needs to be considered. Replacing average drivers either an average system yields no benefit. If it's applied to both tails equally, then you haven't moved the average either, juar made it more homogeneous. You'd have to target the highest risk groups to make the impact. Many in those groups can't afford or won't adopt the tech - the elderly, low income (or less educated depending on the study), or young males mostly.

The better ways to target only the problematic tail is stricter testing, better education, and possibly permemant interlock devices for DUI offenders.

"How many accidents happen because an individual careful driver is, momentaritly, inanttentive?"

My guess would be probably about as often as autonomous systems screw up in novel environments.

> are operating very conservatively right now

If other drivers would not yield to robocar, it may get stuck on intersection forever. Robocars have no assertiveness to push themselves into traffic.

Although I have limited expertise in Boston, there, anyway, "pushing" is part of the driving protocol and expected.

> been issues with numbers for Tesla autopilot

Not surprising.

> To scale up, they would need to drive less conservatively

Can you quantify this? I think that self-driving cars should drive conservatively. Human drivers who are aggressive or risk taking are idiots even if they get to where they are going on time.

Driverless cars should drive like public transport buses (at least around here) / school buses. Conservative, slower than you would like and safe.

They drive more conservatively than the most conservative human driver, mainly because they don’t try at all to cooperate with other vehicles. It’s clearly due to technical reasons, not because they’re being more safe.

For example, if they need to make a lane change for a turn on a congested street, they will wait for someone to let them in rather than “pushing” their way in like a human driver (especially a bus driver) would. When that inevitably doesn’t happen, they are forced to go around the block and try again.

That exact behavior happened during the “Uber” ride I took on a Waymo, and it is possible they have improved since then.

It depends how conservative you mean by conservative.

If you compare a traffic light intersection where there is a clear moment at which you should stop or continue with a roundabout where you need to make a judgement call; being overly conservative would mean stopping at the roundabout until there is 0 traffic.

There was a discussion on this in ep 454 of Freakonomics radio.

> self-driving cars should drive conservatively

I agree, but the specifics get tricky. It is possible to drive so conservatively that it’s unpredictable and therefore dangerous.

I’ve known some terrible conservative drivers in my time; I was a passenger with someone who stopped abruptly in an intersection despite having a green light because they were afraid someone else was going to run the red (while it’s fair to say the person who rear-ended us was at fault, it was an accident that would not happen with a more normal driver).

Maybe I’m nit picking on the difference between conservative and unsurprising.

Yes, conservative should probably mean unsurprising to some degree. Sudden unexpected stops is not truly conservative. Yeah, I can see how the definition is complex.

In the same vein, the average car on US roads is 12 years old. So the average driver is dealing with the safety considerations of a vehicle that isn't a perfectly maintained current model.

Maybe making the point of comparison with professional drivers in fleet cars would be more accurate, but I doubt that data is available.

> If it could scale up and still operate the same way...

Once they scale up, they will be encountering more other Waymo cars on the roads, eliminating a lot of the accidents where someone else crashed into Waymo, or made an upredictable move. So then they could afford to drive less conservatively.

We are pretty sure that our roads would be much safer if all cars were self-driving, but that cannot happen overnight and there will be an interim phase when robo-cars have to mix with human drivers, which is the hard engineering challenge.

>for the types of situations where autopilot was used, human drivers actually had fewer accidents on average.

Do you have any source for that? As far as I'm aware, there simply isn't enough public data released to actually make an unbiased comparison here. So whichever side of this you're on, you can always twist the data to confirm your point of view. Only when Tesla and all other carmakers begin to hand out all of their data, we will get a useful comparison.